TL;DR
Anthropic released Claude Fable 5 on June 9, 2026, as its most capable generally available Claude model and kept a less restricted version, Mythos 5, limited to selected partners. The company says Fable routes flagged cyber, biology, chemistry and distillation requests to Claude Opus 4.8; reviewers at Every called it the strongest coding model they have tested.
Anthropic released Claude Fable 5 on June 9, 2026, calling it its most capable generally available model, while keeping a less restricted twin, Claude Mythos 5, limited to selected partners. The release matters because Anthropic is putting a higher-tier model into public use through a safety system that routes flagged requests to Claude Opus 4.8 instead of giving every public user the model’s full capability on risky topics.
Fable 5 is the safeguarded public version of the same underlying model Anthropic calls Mythos 5, according to the launch materials. The public API name is claude-fable-5. Mythos 5 has some safety limits lifted and is being deployed first through Project Glasswing, Anthropic’s cyber-defense program with the U.S. government, with select biology researchers expected to follow.
Anthropic says Fable 5 uses classifiers watching for risky cybersecurity, biology, chemistry and model-distillation requests. Clear requests, described in the material as more than 95% of use, are answered by Fable 5; flagged requests, described as under 5%, are routed to Claude Opus 4.8. The company says users are told when routing happens, and that the classifiers are tuned conservatively enough to catch some benign requests.
The performance evidence remains sourced to Anthropic and Every. Every, which had a week of early access, said Fable 5 scored 91 out of 100 on its Senior Engineer benchmark, compared with 63 for Opus 4.8 and 62 for GPT-5.5. Anthropic cited a Stripe codebase migration across 50 million lines of Ruby that it said was completed in one day instead of about two months by a team, along with reported gains in drug design, genomics and vision-based coding tasks.
Power Users Get Public Access
The launch gives developers and companies public access to a model class Anthropic places above Opus, while keeping a separate channel for work it treats as higher risk. For teams using AI coding agents, the news is less about a chat upgrade than about longer-running delegation: the sources describe Fable 5 as strongest on large, bounded jobs that can be reviewed after multi-hour runs.
Cost and supply are also part of the change. Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, according to the launch material. That is about twice Opus 4.8 but less than half the earlier Mythos Preview price. Anthropic is also making it free on Pro, Max, Team and Enterprise through June 22 before moving to usage credits, a staged rollout that indicates access will be managed as demand develops.
As an affiliate, we earn on qualifying purchases.
A New Tier Above Opus
Until this release, Anthropic’s top public Claude models sat in the Opus line. Fable 5 and Mythos 5 are described as the first public-facing entries in a Mythos-class tier above Opus, with the distinction between the two names based on policy settings rather than a different base model.
The naming reflects that split. Anthropic’s material links Fable to the Latin fabula and Mythos to Greek mythos, both tied to the idea of a story or account. In product terms, Fable is the version offered broadly with active safety routing, while Mythos remains limited to selected partners for areas where Anthropic is allowing fewer safeguards.
“the most capable model it has ever made generally available”
— Anthropic launch announcement
Classifier Boundaries Need Real Tests
It is not yet clear how often legitimate security, science or model-evaluation work will be routed away from Fable 5 in regular use, how users can challenge a route decision, or how much capability they lose when a request is handled by Opus 4.8. Anthropic says users are notified when routing occurs, but the available material does not give a detailed breakdown by domain, user tier or geography.
Several performance claims also need wider outside testing. Every’s benchmark and Anthropic’s customer examples point to large gains, but independent replication across more codebases, scientific tasks and enterprise workflows is still developing. Access rules for Mythos 5 and data-handling terms for sensitive prompts also remain key questions for regulated teams.
Free Access Ends June 22
Public users can try Fable 5 during the free access window through June 22, 2026, after which Anthropic says usage credits will apply before standard terms arrive later. Developers and enterprise buyers will be watching for clearer routing logs, stable pricing, service limits and more outside benchmark results.
The next test for Anthropic is whether this fallback architecture can let a more capable model reach broad use without creating unacceptable risk or too much friction for legitimate expert work.
Key Questions
What did Anthropic release?
Anthropic released Claude Fable 5, which it describes as its most capable generally available model, and introduced Claude Mythos 5 as a less restricted partner version of the same underlying model.
Are Fable 5 and Mythos 5 different models?
Anthropic’s launch materials describe them as the same underlying model. Fable 5 is the public version with safety routing active, while Mythos 5 has some limits lifted for selected partners.
Does Fable 5 refuse risky prompts?
Anthropic says flagged requests are routed to Claude Opus 4.8, and users are told when this occurs. The system may also catch some benign requests because the classifiers are tuned conservatively.
How good is Fable 5 at coding?
Every said Fable 5 scored 91 out of 100 on its Senior Engineer benchmark, ahead of Opus 4.8 and GPT-5.5 in its testing. That is an outside review claim, not yet a broad industry consensus.
How much does Fable 5 cost?
Launch material lists pricing at $10 per million input tokens and $50 per million output tokens. Anthropic is offering free access on Pro, Max, Team and Enterprise through June 22, 2026, before moving to usage credits.
Source: Thorsten Meyer AI
